Real-Time Brachial Plexus Ultrasound Segmentation Using Lightweight Hierarchical Temporal Fusion

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Anesthesiology

Background:

• Ultrasound-guided brachial plexus blocks are challenging due to difficulties in identifying the brachial plexus, which is operator-dependent.
• Real-time segmentation of nerve elements using AI for these blocks is an understudied area.

Purpose of the Study:

• To introduce a lightweight video segmentation framework for sequential ultrasound imaging in supraclavicular blocks.
• To improve the accuracy and efficiency of brachial plexus identification during ultrasound-guided procedures.

Main Methods:

• A semi-automated pipeline using a tracking algorithm and UltraSam converts bounding boxes to masks, reducing annotation workload.
• A hierarchical temporal fusion module with convolutional recurrent units was integrated into a standard segmentation backbone.
• Training utilized truncated backpropagation through time and domain-specific losses for temporal consistency and spatial compactness.

Main Results:

• The SegFormer-B0-based model showed improved mean IoU (43.32%) and F1 score (56.67%) over a non-temporal baseline with minimal increase in GFLOPs and parameters.
• The autonomous model achieved 4x higher mean IoU and a 3.6x lower temporal instability score compared to SAM2.
• The model demonstrated consistent gains after fine-tuning on cross-vendor devices and achieved real-time performance on mobile NPUs.

Conclusions:

• The developed AI framework significantly enhances real-time ultrasound nerve segmentation for brachial plexus blocks.
• The findings support the potential for safer ultrasound-guided regional anesthesia.
• The framework is adaptable for other real-time medical image segmentation applications.